1 /* 2 * Copyright 2010, Haiku, Inc. All Rights Reserved. 3 * Distributed under the terms of the MIT License. 4 * 5 * Taken from http://burtleburtle.net/bob/c/lookup3.c 6 */ 7 #ifndef LIBKERN_JENKINS_H 8 #define LIBKERN_JENKINS_H 9 10 11 /* 12 ------------------------------------------------------------------------------- 13 lookup3.c, by Bob Jenkins, May 2006, Public Domain. 14 15 These are functions for producing 32-bit hashes for hash table lookup. 16 hashword(), hashlittle(), hashlittle2(), hashbig(), mix(), and final() 17 are externally useful functions. Routines to test the hash are included 18 if SELF_TEST is defined. You can use this free for any purpose. It's in 19 the public domain. It has no warranty. 20 21 You probably want to use hashlittle(). hashlittle() and hashbig() 22 hash byte arrays. hashlittle() is is faster than hashbig() on 23 little-endian machines. Intel and AMD are little-endian machines. 24 On second thought, you probably want hashlittle2(), which is identical to 25 hashlittle() except it returns two 32-bit hashes for the price of one. 26 You could implement hashbig2() if you wanted but I haven't bothered here. 27 28 If you want to find a hash of, say, exactly 7 integers, do 29 a = i1; b = i2; c = i3; 30 mix(a,b,c); 31 a += i4; b += i5; c += i6; 32 mix(a,b,c); 33 a += i7; 34 final(a,b,c); 35 then use c as the hash value. If you have a variable length array of 36 4-byte integers to hash, use hashword(). If you have a byte array (like 37 a character string), use hashlittle(). If you have several byte arrays, or 38 a mix of things, see the comments above hashlittle(). 39 40 Why is this so big? I read 12 bytes at a time into 3 4-byte integers, 41 then mix those integers. This is fast (you can do a lot more thorough 42 mixing with 12*3 instructions on 3 integers than you can with 3 instructions 43 on 1 byte), but shoehorning those bytes into integers efficiently is messy. 44 ------------------------------------------------------------------------------- 45 */ 46 47 #define rot(x,k) (((x)<<(k)) | ((x)>>(32-(k)))) 48 49 /* 50 ------------------------------------------------------------------------------- 51 mix -- mix 3 32-bit values reversibly. 52 53 This is reversible, so any information in (a,b,c) before mix() is 54 still in (a,b,c) after mix(). 55 56 If four pairs of (a,b,c) inputs are run through mix(), or through 57 mix() in reverse, there are at least 32 bits of the output that 58 are sometimes the same for one pair and different for another pair. 59 This was tested for: 60 * pairs that differed by one bit, by two bits, in any combination 61 of top bits of (a,b,c), or in any combination of bottom bits of 62 (a,b,c). 63 * "differ" is defined as +, -, ^, or ~^. For + and -, I transformed 64 the output delta to a Gray code (a^(a>>1)) so a string of 1's (as 65 is commonly produced by subtraction) look like a single 1-bit 66 difference. 67 * the base values were pseudorandom, all zero but one bit set, or 68 all zero plus a counter that starts at zero. 69 70 Some k values for my "a-=c; a^=rot(c,k); c+=b;" arrangement that 71 satisfy this are 72 4 6 8 16 19 4 73 9 15 3 18 27 15 74 14 9 3 7 17 3 75 Well, "9 15 3 18 27 15" didn't quite get 32 bits diffing 76 for "differ" defined as + with a one-bit base and a two-bit delta. I 77 used http://burtleburtle.net/bob/hash/avalanche.html to choose 78 the operations, constants, and arrangements of the variables. 79 80 This does not achieve avalanche. There are input bits of (a,b,c) 81 that fail to affect some output bits of (a,b,c), especially of a. The 82 most thoroughly mixed value is c, but it doesn't really even achieve 83 avalanche in c. 84 85 This allows some parallelism. Read-after-writes are good at doubling 86 the number of bits affected, so the goal of mixing pulls in the opposite 87 direction as the goal of parallelism. I did what I could. Rotates 88 seem to cost as much as shifts on every machine I could lay my hands 89 on, and rotates are much kinder to the top and bottom bits, so I used 90 rotates. 91 ------------------------------------------------------------------------------- 92 */ 93 #define mix(a,b,c) \ 94 { \ 95 a -= c; a ^= rot(c, 4); c += b; \ 96 b -= a; b ^= rot(a, 6); a += c; \ 97 c -= b; c ^= rot(b, 8); b += a; \ 98 a -= c; a ^= rot(c,16); c += b; \ 99 b -= a; b ^= rot(a,19); a += c; \ 100 c -= b; c ^= rot(b, 4); b += a; \ 101 } 102 103 /* 104 ------------------------------------------------------------------------------- 105 final -- final mixing of 3 32-bit values (a,b,c) into c 106 107 Pairs of (a,b,c) values differing in only a few bits will usually 108 produce values of c that look totally different. This was tested for 109 * pairs that differed by one bit, by two bits, in any combination 110 of top bits of (a,b,c), or in any combination of bottom bits of 111 (a,b,c). 112 * "differ" is defined as +, -, ^, or ~^. For + and -, I transformed 113 the output delta to a Gray code (a^(a>>1)) so a string of 1's (as 114 is commonly produced by subtraction) look like a single 1-bit 115 difference. 116 * the base values were pseudorandom, all zero but one bit set, or 117 all zero plus a counter that starts at zero. 118 119 These constants passed: 120 14 11 25 16 4 14 24 121 12 14 25 16 4 14 24 122 and these came close: 123 4 8 15 26 3 22 24 124 10 8 15 26 3 22 24 125 11 8 15 26 3 22 24 126 ------------------------------------------------------------------------------- 127 */ 128 #define final(a,b,c) \ 129 { \ 130 c ^= b; c -= rot(b,14); \ 131 a ^= c; a -= rot(c,11); \ 132 b ^= a; b -= rot(a,25); \ 133 c ^= b; c -= rot(b,16); \ 134 a ^= c; a -= rot(c,4); \ 135 b ^= a; b -= rot(a,14); \ 136 c ^= b; c -= rot(b,24); \ 137 } 138 139 /* 140 -------------------------------------------------------------------- 141 This works on all machines. To be useful, it requires 142 -- that the key be an array of uint32's, and 143 -- that the length be the number of uint32's in the key 144 145 The function hashword() is identical to hashlittle() on little-endian 146 machines, and identical to hashbig() on big-endian machines, 147 except that the length has to be measured in uint32s rather than in 148 bytes. hashlittle() is more complicated than hashword() only because 149 hashlittle() has to dance around fitting the key bytes into registers. 150 -------------------------------------------------------------------- 151 */ 152 static uint32 153 jenkins_hashword(const uint32 *k, /* the key, an array of uint32 values */ 154 size_t length, /* the length of the key, in uint32s */ 155 uint32 initval) /* the previous hash, or an arbitrary value */ 156 { 157 uint32 a,b,c; 158 159 /* Set up the internal state */ 160 a = b = c = 0xdeadbeef + (((uint32)length)<<2) + initval; 161 162 /*------------------------------------------------- handle most of the key */ 163 while (length > 3) 164 { 165 a += k[0]; 166 b += k[1]; 167 c += k[2]; 168 mix(a,b,c); 169 length -= 3; 170 k += 3; 171 } 172 173 /*------------------------------------------- handle the last 3 uint32's */ 174 switch(length) /* all the case statements fall through */ 175 { 176 case 3 : c+=k[2]; 177 case 2 : b+=k[1]; 178 case 1 : a+=k[0]; 179 final(a,b,c); 180 case 0: /* case 0: nothing left to add */ 181 break; 182 } 183 /*------------------------------------------------------ report the result */ 184 return c; 185 } 186 187 #endif 188